1 / 0

A pre-Weekend Talk on Online Learning

A pre-Weekend Talk on Online Learning. TGIF Talk Series Purushottam Kar. Outline. Some Motivating Examples Discovering customer preferences Learning investment profiles Detecting credit card fraud The Formal Online Learning Framework Notion of regret Formalization of motivating examples

dolan
Download Presentation

A pre-Weekend Talk on Online Learning

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A pre-Weekend Talk on Online Learning

    TGIF Talk Series Purushottam Kar
  2. Outline Some Motivating Examples Discovering customer preferences Learning investment profiles Detecting credit card fraud The Formal Online Learning Framework Notion of regret Formalization of motivating examples Simple Online Algorithms Online classification, regression Online ranking Batch solvers for large scale learning problems Other “Feedback-based” Learning Frameworks
  3. Some Motivating Examples Why Online Learning can be Useful
  4. The Cook’s Dilemma
  5. Discovering Customer Preferences Loss 2 1 1 0 2
  6. Learning Investment Profiles assets that give returns proportional to investment Asset gives back as return per dollar invested If I invest in then total return is Return profile depends on market forces, other investors and keeps changing I have corpus of that I decide to invest completely in these assets Let decide proportion of investment in asset , i.e. investment is Corpus at time becomes : reward to be maximized
  7. Detecting Credit Card Fraud Classify credit card payments into Each payment is described by a vector Other problems such as branch prediction/churn prediction Linear classification model Choose and classify as Online process; at each time A credit card payment is detected We propose a linear classifier and classify as True status of payment is made known and our mistake (if any) is revealed Wish to minimize the number of mistakes made by us Wish to propose a “good” sequence of
  8. The Formal Online Learning Framework How we assess Online Learning Algorithms
  9. The Online Learning Model An attempt to model an interactive and adaptive environment We have a set of actions Environment has a set of loss functions In each round We play some action Environment responds with a loss function We are forced to incur a loss Environment can adapt to our actions (or even be adversarial) Our goal: minimize cumulative loss Can cumulative loss be brought down to zero : mostly no ! More reasonable measure of performance: single best action in hindsight Regret: Why is this a suitable notion of performance ?
  10. Motivating Examples Revisited Detecting customer preferences Assume we can represent customer as a vector Set of actions are linear functions predicting spice levels for that customer Loss function given by squared difference between true and preferred spiciness At time step customer comes and Goal: make customers as happy as the single best spice level Credit card fraud detection Actions are the set of linear classifiers Loss functions are mistake functions Detection of credit card fraud might change buying profiles (adversarial) Goal: make (almost) as few mistakes as single best classifier
  11. Motivating Examples Revisited Learning investment profiles Set of actions is the -dimensional simplex Reward received at step is where is the return given by market Total reward (assume w.l.o.g. initial corpus is ) Returns affected by investment, other market factors (adaptive, adversarial) Can think of as a negative reward or a loss Regret (equivalently) given by Goal: make as much profit as the single best investment profile in hindsight
  12. Simple Online Algorithms What makes online learning click ?
  13. Online Linear Classification Perceptron Algorithm Start with Classify as If correct classification i.e. , then let Else Loss function i.e. 1 iff misclassifies If there exists a perfect linear separator such that , If there exists an imperfect separator such that ,
  14. The Perceptron Algorithm in action
  15. Online Regression The Perceptron Algorithm was (almost) a gradient descent algorithm Consider the loss function is a convex surrogate to the mistake function When perceptron makes a mistake i.e. , we have Thus the perceptron update step is a gradient step !
  16. Online Regression via Online Gradient Descent Suppose we are taking actions and receiving losses Assume that all loss function are convex and Lipchitz Examples , , Online Gradient Descent (for linear predictions problems) Start with Receive object and predict value for object Receive loss function and update Some more work needed to ensure that as well We can ensure that
  17. Online Bipartite Ranking Documents arrive in a continuous stream to be ranked Each document in labelled either “relevant” (+) or “irrelevant” (-) Goal: somehow rank all relevant documents before irrelevant ones Method: assign relevance score to document and sort We incur loss for “swaps” Minimize number of swaps Problem is equivalent to maximizing area under the ROC curve of TP/FP Challenges No reference point: no “true” relevance score Need pairs of documents to learn a scoring function: get only singles Solution: keep (some) of the past points in a buffer to construct pairs on the fly Several interesting algorithmic and theoretical problems still open
  18. Batch Solvers Statistical learning gets a batch of randomly chosen training examples We wish to learn a function that does well on these exampleswhere is a loss function (classification, regression etc) Statistical Learning Theory: such an does well on unseen points as well! Solving “batch” problem may be infeasible: , distributed storage etc. Solution: solve the online problem instead E.g. online gradient descent will solve for a such that where
  19. Batch Solvers Thus we have an such thatwhere Online to batch conversion bounds Argue for the performance of on random unseen points Expected loss of on a random unseen point is bounded Several batch solvers e.g. PEGASOS, MIDAS, LASVM use techniques such as Stochastic online gradient descent for large scale learning
  20. Other Feedback based Learning Frameworks Two axes of variation: modelling of environment and feedback Online Learning: some modelling of environment and full feedback Losses are simple functions over linear models (can be made non linear though) At each step the loss function itself is given to us: full information Models are agnostic: no realizability assumptions are made Multi-armed Bandits: weak modelling of environment, weak feedback Often no assumptions made on nature of loss function Limited feedback: only loss value on played action made available Contextual bandits try to model loss function but make realizability assumptions Reinforcement Learning: Strong modelling of environment, weak feedback Environment modelled as a state space with adaptive stochastic transitions Reward functions modeled as functions of state space and action Limited feedback available: need to learn, state space as well as reward function
More Related